The Assumption Operators Need to Pressure-Test
Most teams building AI-assisted GTM pipelines treat agent count as a proxy for capability. More agents, more specialization, better output. It's an intuitive model — and Crawford's live experiment suggests it's often wrong.
In a documented compression test, Crawford reduced a large multi-agent system to roughly 1/30th of its original agent count and measured output agreement against the original system. The result: 87% agreement. Not 50%. Not 70%. Eighty-seven percent — from a stack that's a fraction of the compute, coordination overhead, and maintenance surface area.
For non-tech operators standing up first-generation outbound or enrichment pipelines, this is one of the few empirical data points available on agent efficiency. Most GTM agent advice is architectural opinion. This is a ratio you can benchmark against.
Why Coordination Overhead Is the Hidden Cost
The problem with multi-agent bloat isn't just compute cost. It's the coordination tax: handoffs that introduce latency, error propagation between agents, prompt chains that are hard to debug, and systems that require specialist attention to maintain. For a company without a dedicated AI engineering team, that tax compounds fast.
When you collapse agents, you also collapse failure modes. A leaner system is easier to audit, easier to fix when something drifts, and easier to hand off to an operator who isn't a prompt engineer. At the margin, a system your RevOps lead can actually maintain beats an optimized system only your implementation vendor understands.
The 87% agreement figure also reframes what "good enough" looks like in GTM automation. If 13% of output variation is acceptable — and in most enrichment or outbound personalization contexts it is — then the question isn't "how do I get to 100%?" It's "how lean can I run while staying above my quality threshold?"
The Compression Test You Can Run This Week
If you have an existing multi-agent GTM workflow — even a modest one — here's the practical move:
- Document your current agent chain and what each agent is responsible for.
- Identify agents with overlapping or sequential similar tasks — these are your first consolidation candidates.
- Collapse two to three agents into one with a combined prompt, run it against 50–100 records, and score output agreement against your original system.
- Set your agreement threshold before you run it. For most B2B enrichment or outbound copy use cases, 80%+ is a defensible bar.
If you're still in the design phase, start with the minimum number of agents that can produce a shippable output. You can always add specialization later. You rarely need to.
The goal isn't minimalism for its own sake. It's building a system your team can operate, validate, and improve — without an engineering team behind it.